Intel CEO: CPU only meets 50% of demand, 14A to start production in Q1 next year, new architecture may reduce inference power consumption to 1/15 of GPU

On September 14, Intel CEO Pat Gelsinger stated in a recent fireside chat at an industry summit that as AI agents scale rapidly, CPU demand has far exceeded Intel’s current supply capabilities, and at present, the company can only meet about half of the demand from leading clients.

Meanwhile, he revealed that the company’s process technology has made significant advancements: the 18A node has entered mass production and the 14A node will begin production in the first quarter of next year.

Gelsinger stated that multiple large tech company CEOs have directly called to request more CPU supply, “I had to apologize because our capacity couldn’t keep up.”

He attributed this shortage to the surging CPU demand in AI inference scenarios. CPUs hold irreplaceable advantages over GPUs when it comes to agent orchestration, control planes, and multi-threaded task processing.

Additionally, Gelsinger revealed that Intel is supporting new types of dedicated chips based on dataflow architecture and wafer-level scaling solutions, which can deliver the same performance as GPUs at only 1/10 to 1/15 the power consumption in specific inference scenarios.

On the market capitalization front, Gelsinger mentioned that when he took over as CEO, Intel’s market value was about $90 billion, which has now risen to approximately $500 billion.

On yield improvement in the manufacturing process, Gelsinger disclosed that by opening up factory data to equipment manufacturers and data analytics partners, Intel has achieved an annual yield increase of about 7%. He admitted that factory yields were “extremely poor” when he first took office, but emphasized that breaking down silos and seeking external help was key to turning things around.

The Era of AI Agents Drives a Compute Power Shift as High-End CPUs Face Severe Shortages

There is a widespread misconception in the market that AI development relies solely on GPUs.

Gelsinger noted that although GPUs excel in model training, compute requirements are reversing once AI systems reach the agent and reinforcement learning inference stages. He stated:

When orchestration, control flows, and complex single-threaded or multi-threaded scheduling are needed, CPUs are the best choice.

Gelsinger pointed out that leading edge model vendors are all scrambling to purchase high-end CPUs, directly resulting in extremely tight Intel capacity.

He admitted that current delivery pressures are extremely high:

I can only provide 50% of the CPU orders my customers need. Many tech giants’ CEOs are calling me, lining up to buy, and I have to apologize to them because I’m just not making enough.

He clearly predicted that as millions or even trillions of AI agents go online in the future, the compute infrastructure including CPUs, memory, and network bandwidth will continue to be “supply constrained.”

Recalling the challenges at the start of his tenure as CEO, Gelsinger stated that the depth and breadth of the company’s problems far exceeded what he had inferred during his two years as a board member—“at least ten times more than I expected.”

He listed cultural transformation as his top priority: first, to establish a culture of accountability; second, to truly listen to customers.

He revealed that soon after he took office, a major client laid out 15 mistakes Intel had made,

They told me, your teams don’t listen to us and only lecture us, which is why we designed you out.

Gelsinger also emphasized that engineering teams report directly to him, allowing him to access true frontline information, rather than information “airbrushed” through layers of management.

He sees maintaining continuous connections with the startup and venture capital community as a key mechanism to ensure Intel never misses the next tech wave again.

18A Fully in Mass Production, 14A to Enter Production in Q1 Next Year

Wafer foundry services are at the core of Intel’s revival, yet Gelsinger stressed that foundry is essentially a “customer service business,” which requires a complete overhaul of Intel’s past culture. Gelsinger said:

Foundry isn’t about being aloof—you have to serve customers with delight and humility, seamlessly supporting their preferred EDA tools or third-party IP libraries.

To address the yield issues of advanced processes, Gelsinger took decisive action to break the entrenched internal practice of engineers hiding data. He recalled this impactful process:

I had to change the culture and fired some engineers to make sure we were totally transparent. When we opened up the real data, CEO friends from two equipment companies called me and said: “Pat, I have to tell you some bad news, this initial yield is terrible.” I replied: “That’s exactly why I’m asking for your help.”

With joint efforts, Intel’s advanced process technology made a huge leap forward. Gelsinger officially confirmed the latest timeline and data highlights:

The good news is, now I’m seeing a 7% improvement in yield every year. Intel’s 18A process is already in mass production, and Intel’s 14A process will officially start production in the first quarter of next year.

Advanced Packaging Is the “Holy Grail”—Massive US Investments to Address Supply Chain Risks

With Moore’s Law slowing, Gelsinger considers advanced packaging the industry’s “ultimate holy grail” for the next five years.

Gelsinger believes that system-level packaging, integrating CPUs, large memory, and I/O, is key to increasing compute density.

However, he also issued a stark warning about the supply chain. He pointed out that global key advanced substrate materials are monopolized by only a handful of vendors (two in Japan and two in the Taiwan region of China), who require enormous prepayments for production capacity. Gelsinger said:

95% of the world’s advanced packaging capacity is concentrated in a single region, which is extremely risky.

Intel is making massive capital expenditures to accelerate building advanced packaging and manufacturing capabilities in the United States.

“10-Year Dimensionality Reduction Plan” and Quantum Computing Node

Regarding current AI infrastructure development, Gelsinger voiced deep concerns about power consumption.

Jensen Huang has done an excellent job in training, but the entire industry is facing an insurmountable problem—power consumption limits.

To break this bottleneck, Gelsinger revealed the “decade-long forward-looking project” he is secretly working on.

He pointed out that the human brain handles extremely complex cognition with only tens of watts of energy—showing that a “10,000-fold reduction in power consumption” is potentially viable for future compute power.

Currently, Intel is supporting startups (such as SambaNova, Cerebras) that are based on dataflow architecture and wafer-level scaling solutions, “which can deliver the same performance as GPUs at only 1/10 to 1/15 the power consumption in specific inference scenarios.”

At the same time, the company is heavily investing in Neuromorphic Computing and brain-like MPUs.

For the cutting edge of quantum computing, Gelsinger gave a clear commercialization timeline. He stated that the company is now highly focused on solving quantum error correction:

It is expected that quantum computing will make a substantial industrial impact within the next three to five years, forming a heterogeneous computing network with CPUs and GPUs to jointly tackle high-dimensional computational challenges.